pith. sign in

arxiv: 2606.10023 · v1 · pith:CKTNI6NFnew · submitted 2026-06-08 · 🌌 astro-ph.CO · astro-ph.IM· cs.LG

Learning the Universe: Posterior Reliability of Neural Generative Models in High-Dimensional Field-Level Inference of Cosmic Initial Conditions

Ludvig Doeser , Jens Jasche This is my paper

Pith reviewed 2026-06-27 15:20 UTC · model grok-4.3

classification 🌌 astro-ph.CO astro-ph.IMcs.LG
keywords cosmological inferencefield-level inferencegenerative modelsposterior reliabilityinitial conditionsHamiltonian Monte Carlouncertainty quantificationnormalizing flows
0
0 comments X

The pith

Matching posterior means and correlations does not ensure correct uncertainty structure in neural generative models for cosmic initial conditions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper evaluates two neural generative models for inferring high-dimensional cosmic initial conditions from present-day large-scale structure. It uses Hamiltonian Monte Carlo samples as a reference to test an implicit stochastic interpolants model and an explicit GLOW normalizing flow model. Standard checks such as agreement on posterior means, marginal distributions, and cross-correlations are shown to miss failures in the posterior geometry. These failures appear clearly in comparisons of posterior variance fields and in sample-based evaluations. The result matters because reliable uncertainty estimates determine what can be learned from cosmological observations when using fast amortized inference methods.

Core claim

In the cosmological inverse problem of recovering initial conditions, both the stochastic interpolants model and the GLOW model reproduce posterior means and achieve high cross-correlations with reference samples, yet produce posterior variance fields that differ from the Hamiltonian Monte Carlo reference and fail sample-based tests of uncertainty structure.

What carries the argument

Posterior variance fields and sample-based evaluations, which expose geometry failures in the inferred posteriors that are invisible to mean-matching or correlation metrics.

If this is right

  • Agreement on posterior means, marginal distributions, or cross-correlations alone is insufficient to confirm that a generative model has recovered the correct uncertainty structure.
  • Field-level inference with neural generative models requires validation against reference samples that probe the full posterior geometry, not just summary statistics.
  • Generative models can handle non-differentiable simulators in cosmology, but only after explicit checks confirm that uncertainty estimates remain reliable.
  • Standard metrics commonly used to assess amortized inference can give misleading signals of success in high-dimensional field problems.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same variance-field diagnostic could be applied to test posterior reliability in other high-dimensional inverse problems that use amortized generative models.
  • If variance mismatches prove systematic, it may restrict the precision of downstream cosmological parameter constraints derived from these models.
  • Extending the evaluation to include higher-order statistics or information content metrics could further clarify when generative models preserve the necessary posterior structure.

Load-bearing premise

Hamiltonian Monte Carlo produces sufficiently accurate reference posterior samples to serve as ground truth for testing the generative models.

What would settle it

A direct side-by-side comparison in which the generative models' posterior variance fields match the Hamiltonian Monte Carlo variance fields across the simulated volume while means and correlations also agree.

Figures

Figures reproduced from arXiv: 2606.10023 by Jens Jasche, Ludvig Doeser.

Figure 1
Figure 1. Figure 1: Voxel-wise posterior mean and variance in 2D slices of the 3D final present-day density field (top) and the initial conditions field (bottom) in a volume of 1ℎ −1 Gpc and ∼ 30ℎ −1 Mpc resolution per voxel. Results are estimated from samples drawn using HMC, a conditional GLOW model, and SI model trained on the rnd data (see [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Same as [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Marginalized posterior distributions for representative voxels spanning high-, intermediate-, and low-density environments. The mock data (dashed red) correspond to the true final field (dashed black) with added Gaussian noise. All methods yield consistent posteriors and typically recover values closer to the underlying truth, demonstrating that voxel-wise noise does not prevent them from exploiting large-… view at source ↗
Figure 4
Figure 4. Figure 4: Voxel-wise posterior variance over samples of the reconstructed final (top) and initial (bottom) fields compared to the HMC posterior variance, where each point corresponds to a voxel. The dashed line indicates perfect agreement, while colours denote the posterior mean density, illustrating that the spatial structure of the variance of the final field closely follows the underlying density field. The repor… view at source ↗
Figure 5
Figure 5. Figure 5: Scale-dependent cross-correlation between reconstructed and true fields for the initial (left) and final fields (right). The top row correlates samples from HMC, GLOW, and SI to the truth, while the bottom row shows mean fields of GLOW/SI relative to the HMC mean. 0.8 1.0 1.2 Tk,mean Initial field Final field 10−2 10−1 k[hMpc−1 ] 0.8 1.0 1.2 Tk,variance 10−2 10−1 k[hMpc−1 ] [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 6
Figure 6. Figure 6: Power-spectrum ratios relative to HMC. Top row: the ratio for the posterior mean, which probes the recovery of spatial structure across scales. Bottom row: the ratio for the variance, quantifying spatial correlations in posterior uncertainty across scales. Colors match [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Histograms of unnormalized log prior, log likelihood, and log posterior evaluations, all divided by the dimension 𝑑 = 323 to aid visualization and interpretation. Thus, the values shown correspond to contributions per voxel (e.g. the Gaussian prior contributes −1/2 per dimension). Distributions of samples generated by each method reflect how effectively each sampler ultimately captures regions of the targe… view at source ↗
Figure 8
Figure 8. Figure 8: Posterior ratios (PR) of GLOW-predicted probabilities and true (Eq. (12)) log-posteriors for pairs of realizations conditioned on the same data. Each point corresponds to an independent random pair of samples, drawn either from GLOW (left) or from HMC (right). The corresponding lines are linear fits to each model. The bottom panels show the residuals from the dashed lines, which indicate perfect agreement,… view at source ↗
Figure 9
Figure 9. Figure 9: Scale-dependent cross-correlation between reconstructed and true fields as a function of the number of training simulations for the GLOW and SI models. The top row shows the final fields and the bottom row the initial conditions. Increasing the number of training simulations yields modest improvements in the recovery of structures across scales, an effect that is slightly more evident in the final field. F… view at source ↗
Figure 10
Figure 10. Figure 10: Performance as a function of the number of training samples, evaluated through the unnormalized log-prior (top), log-likelihood (middle), and log-posterior (bottom). Increasing the training set size systematically improves agreement with the HMC reference posterior for all models. The overall convergence behaviour appears similar across architectures and training strategies. Only GLOWrnd was trained on th… view at source ↗
read the original abstract

Accurate posterior estimation is central to scientific inference, as uncertainties determine what can be reliably learned from observational data. While Markov chain Monte Carlo methods provide asymptotic convergence guarantees, they are computationally demanding in high-dimensional settings. Neural network-based generative models for entire discretized 3D fields enable fast amortized inference but often lack convergence guarantees and principled accuracy assessment. Using Hamiltonian Monte Carlo to obtain reference posterior samples, we conduct a controlled field-level evaluation of an implicit generative model (Stochastic Interpolants) and an explicit likelihood-based model (GLOW normalizing flows). This comparison, unavailable in typical applications, enables the detection of posterior geometry failures that standard metrics cannot capture. As a case study, we consider the cosmological inverse problem of inferring cosmic initial conditions from present-day large-scale structure. To match the precision of modern cosmological data, this problem increasingly relies on complex, non-linear, and non-differentiable simulators, which are incompatible with gradient-based inference frameworks. Generative models offer a route to address these challenges, provided their inferred posteriors are reliable. In this work, we show that matching posterior means, marginal distributions, or achieving high cross-correlation does not imply correct uncertainty structure, as revealed by posterior variance fields and sample-based evaluations. Through this work, we aim to raise awareness of the challenges of uncertainty estimation in high-dimensional field-level settings, highlighting the importance of careful design and validation of neural generative approaches for scientific applications.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript claims that in the high-dimensional field-level inference of cosmic initial conditions from large-scale structure, agreement between neural generative models (Stochastic Interpolants and GLOW normalizing flows) and HMC reference samples on posterior means, marginal distributions, or cross-correlations does not guarantee correct uncertainty structure. This is demonstrated via discrepancies in posterior variance fields and sample-based evaluations, using a controlled comparison enabled by HMC references that is unavailable in typical applications.

Significance. If the central claim holds, the work is significant for highlighting limitations of standard posterior diagnostics in amortized inference for cosmology. It provides a concrete case study showing that conventional metrics can miss geometry failures in generative models, which could inform validation practices for high-dimensional scientific applications where simulators are non-differentiable.

major comments (2)
  1. [Abstract / Evaluation design] The central claim that generative models exhibit failures in uncertainty structure (revealed by variance fields and sample-based evaluations) rests on HMC samples serving as faithful ground truth. However, no convergence diagnostics, effective sample sizes, or cross-validation against an independent sampler are described, which is load-bearing in this high-dimensional, non-Gaussian, non-linear setting.
  2. [Abstract] The abstract states that modern simulators are non-differentiable and thus incompatible with gradient-based inference frameworks, yet HMC (a gradient-based method) is used to generate the reference posterior samples for the controlled comparison. Clarification is required on how gradients are obtained or approximated to support the validity of the reference.
minor comments (1)
  1. [Abstract] The abstract supplies no quantitative results, error bars, dataset sizes, or specific metrics for the claimed discrepancies, which limits immediate assessment of effect sizes.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their careful reading and constructive comments, which help strengthen the manuscript. We address each major comment below and will incorporate clarifications and additional diagnostics in the revised version.

read point-by-point responses

  1. Referee: [Abstract / Evaluation design] The central claim that generative models exhibit failures in uncertainty structure (revealed by variance fields and sample-based evaluations) rests on HMC samples serving as faithful ground truth. However, no convergence diagnostics, effective sample sizes, or cross-validation against an independent sampler are described, which is load-bearing in this high-dimensional, non-Gaussian, non-linear setting.

    Authors: We agree that explicit convergence diagnostics are essential to substantiate the HMC reference samples as reliable ground truth. The revised manuscript will include Gelman-Rubin statistics, effective sample size estimates for field-level and summary statistics, and autocorrelation times. These will be added to the methods and supplementary sections to address this point directly. revision: yes

  2. Referee: [Abstract] The abstract states that modern simulators are non-differentiable and thus incompatible with gradient-based inference frameworks, yet HMC (a gradient-based method) is used to generate the reference posterior samples for the controlled comparison. Clarification is required on how gradients are obtained or approximated to support the validity of the reference.

    Authors: The abstract highlights the general challenge with non-differentiable simulators in modern cosmology. For the controlled experiment, however, we used a differentiable particle-mesh forward model (with adjoint gradients) to enable HMC sampling and obtain reference posteriors. This setup isolates the comparison while the generative models are evaluated in a setting where only non-differentiable simulators would be available in practice. We will revise the abstract and methods to explicitly distinguish the reference model from the general case. revision: yes

Circularity Check

0 steps flagged

No significant circularity; HMC reference samples provide independent ground truth for model evaluation

full rationale

The paper's central claim—that standard posterior metrics (means, marginals, cross-correlations) fail to guarantee correct uncertainty structure—is demonstrated by direct comparison of generative model outputs against separately generated HMC reference samples. No equations reduce any claimed result to quantities defined by the models under test, no parameters are fitted and then relabeled as predictions, and no load-bearing steps rely on self-citations or imported uniqueness theorems. The evaluation chain is self-contained against an external MCMC benchmark whose convergence properties are treated as given rather than derived from the neural models.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review yields no explicit free parameters, axioms, or invented entities beyond the background assumption that HMC supplies reliable references; no further ledger entries can be extracted.

pith-pipeline@v0.9.1-grok · 5795 in / 1053 out tokens · 22299 ms · 2026-06-27T15:20:43.906966+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

130 extracted references · 93 canonical work pages · 29 internal anchors

  1. [1]

    ascl , howpublished =:1811.008 , adsnote =

    Pylians:. ascl , howpublished =:1811.008 , adsnote =

  2. [2]

    Journal of Machine Learning Research , volume =

    Stochastic Interpolants: A Unifying Framework for Flows and Diffusions , author =. Journal of Machine Learning Research , volume =

  3. [3]

    Coverage is not enough: Frequentist tests of simulation-based inference for primordial non-Gaussianity

    Alokda, Toka and Porciani, Cristiano and Eggemeier, Alexander , year = 2026, month = may, number =. Coverage Is Not Enough:. doi:10.48550/arXiv.2605.00980 , urldate =. arXiv , keywords =:2605.00980 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2605.00980 2026

  4. [4]

    Fast likelihood-free cosmology with neural density estimators and active learning

    Fast Likelihood-Free Cosmology with Neural Density Estimators and Active Learning , author =. Monthly Notices of the Royal Astronomical Society , volume =. doi:10.1093/mnras/stz1960 , urldate =. arXiv , keywords =:1903.00007 , pages =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1093/mnras/stz1960 1903

  5. [5]

    Cosmology and Fundamental Physics with the

    Amendola, Luca and Appleby, Stephen and Avgoustidis, Anastasios and Bacon, David and Baker, Tessa and Baldi, Marco and Bartolo, Nicola and Blanchard, Alain and Bonvin, Camille and Borgani, Stefano and Branchini, Enzo and Burrage, Clare and Camera, Stefano and Carbone, Carmelita and Casarini, Luciano and Cropper, Mark and. Cosmology and Fundamental Physics...

    work page doi:10.1007/s41114-017-0010-3

  6. [6]

    Acta Numerica , volume =

    Solving Inverse Problems Using Data-Driven Models , author =. Acta Numerica , volume =. doi:10.1017/S0962492919000059 , urldate =

    work page doi:10.1017/s0962492919000059

  7. [7]

    and Kashino, Daichi and Cucciati, Olga and

    Ata, Metin and Kitaura, Francisco Shu and Lee, Khee Gan and Lemaux, Brian C. and Kashino, Daichi and Cucciati, Olga and. Monthly Notices of the Royal Astronomical Society , volume =. doi:10.1093/mnras/staa3318 , urldate =. arXiv , keywords =:2004.11027 , pages =

    work page doi:10.1093/mnras/staa3318 2004

  8. [8]

    , year = 2024, month = jul, number =

    Augustine, Midhun T. , year = 2024, month = jul, number =. A. doi:10.48550/arXiv.2407.12895 , urldate =. arXiv , keywords =:2407.12895 , primaryclass =

    work page doi:10.48550/arxiv.2407.12895 2024

  9. [9]

    2025, title The BIG SOBOL SEQUENCE: How many simulations do we need for simulation-based inference in cosmology? , , 703, A301, 10.1051/0004-6361/202554602

    How Many Simulations Do We Need for Simulation-Based Inference in Cosmology? , author =. Astronomy & Astrophysics , volume =. doi:10.1051/0004-6361/202554602 , urldate =. arXiv , keywords =:2503.13755 , primaryclass =

    work page doi:10.1051/0004-6361/202554602

  10. [10]

    and Chiarenza, Marco and Doeser, Ludvig and Leclercq, Florent , year = 2025, month = sep, journal =

    Bartlett, Deaglan J. and Chiarenza, Marco and Doeser, Ludvig and Leclercq, Florent , year = 2025, month = sep, journal =. arXiv , file =:2409.02154 , pages =

    arXiv 2025

  11. [11]

    and Seljak, Uros and Modi, Chirag , year = 2023, month = jul, eprint =

    Bayer, Adrian E. and Seljak, Uros and Modi, Chirag , year = 2023, month = jul, eprint =. Field-

    2023

  12. [12]

    and Parker, Liam and Valcin, David and Chen, Shi-Fan and Modi, Chirag and Seljak, Uros , year = 2026, month = mar, number =

    Bayer, Adrian E. and Parker, Liam and Valcin, David and Chen, Shi-Fan and Modi, Chirag and Seljak, Uros , year = 2026, month = mar, number =. Field-. doi:10.48550/arXiv.2603.15732 , urldate =. arXiv , keywords =:2603.15732 , primaryclass =

    work page doi:10.48550/arxiv.2603.15732 2026

  13. [13]

    Benning, Martin and Burger, Martin , year = 2018, month = jan, number =. Modern. doi:10.48550/arXiv.1801.09922 , urldate =. arXiv , keywords =:1801.09922 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1801.09922 2018

  14. [14]

    Large-Scale Structure of the Universe and Cosmological Perturbation Theory

    Bernardeau, F. and Colombi, S. and Gaztanaga, E. and Scoccimarro, R. , year = 2002, month = sep, journal =. Large-. doi:10.1016/S0370-1573(02)00135-7 , urldate =. arXiv , keywords =:astro-ph/0112551 , pages =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1016/s0370-1573(02)00135-7 2002

  15. [15]

    arXiv: 2502.11846 , eprint =

    Neural Network Reconstruction of Non-Gaussian Initial Conditions from Dark Matter Halos , author =. arXiv: 2502.11846 , eprint =

    arXiv

  16. [16]

    Proceedings of the National Academy of Sciences , volume =

    Mining Gold from Implicit Models to Improve Likelihood-Free Inference , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.1915980117 , urldate =. arXiv , keywords =:1805.12244 , primaryclass =

    work page doi:10.1073/pnas.1915980117

  17. [17]

    doi:10.48550/arXiv.2508.02602 , urldate =

    Trustworthy Scientific Inference with Generative Models , author =. doi:10.48550/arXiv.2508.02602 , urldate =. arXiv , keywords =:2508.02602 , primaryclass =

    work page doi:10.48550/arxiv.2508.02602

  18. [18]

    Monthly Notices of the Royal Astronomical Society , volume =

    Effective Cosmic Density Field Reconstruction with Convolutional Neural Network , author =. Monthly Notices of the Royal Astronomical Society , volume =. doi:10.1093/mnras/stad1868 , abstract =. arXiv , keywords =:2306.10538 , pages =

    work page doi:10.1093/mnras/stad1868

  19. [19]

    , year = 2024, journal =

    Chen, Xinyi and Padmanabhan, Nikhil and Eisenstein, Daniel J. , year = 2024, journal =. Probing Primordial Non-. 2412.00968 , archiveprefix =

    arXiv 2024

  20. [20]

    Field-Level

    Chen, Shu-Fan and Chen, Kai-Feng and Dvorkin, Cora , year = 2025, month = nov, journal =. Field-Level. doi:10.1088/1475-7516/2025/11/082 , urldate =. arXiv , keywords =:2508.13265 , primaryclass =

    work page doi:10.1088/1475-7516/2025/11/082 2025

  21. [21]

    Proceedings of the National Academy of Sciences , volume =

    The Frontier of Simulation-Based Inference , author =. Proceedings of the National Academy of Sciences , volume =. doi:10.1073/pnas.1912789117 , urldate =. arXiv , keywords =:1911.01429 , primaryclass =

    work page doi:10.1073/pnas.1912789117 1911

  22. [22]

    Joint Cosmological Parameter Inference and Initial Condition Reconstruction with

  23. [23]

    Mathematics of Control, Signals and Systems , volume =

    Approximation by Superpositions of a Sigmoidal Function , author =. Mathematics of Control, Signals and Systems , volume =. doi:10.1007/BF02551274 , abstract =

    work page doi:10.1007/bf02551274

  24. [24]

    Multiscale

    Dai, Biwei and Seljak, Uros , year = 2024, month = feb, number =. Multiscale. doi:10.48550/arXiv.2306.04689 , urldate =. arXiv , langid =:2306.04689 , primaryclass =

    work page doi:10.48550/arxiv.2306.04689 2024

  25. [25]

    and Macke, Jakob H

    Dax, Maximilian and Wildberger, Jonas and Buchholz, Simon and Green, Stephen R. and Macke, Jakob H. and Sch. Flow. doi:10.48550/arXiv.2305.17161 , urldate =. arXiv , keywords =:2305.17161 , primaryclass =

    work page doi:10.48550/arxiv.2305.17161

  26. [26]

    Deistler, Michael and Boelts, Jan and Steinbach, Peter and Moss, Guy and Moreau, Thomas and Gloeckler, Manuel and Rodrigues, Pedro L. C. and Linhart, Julia and Lappalainen, Janne K. and Miller, Benjamin Kurt and Gon. Simulation-. doi:10.48550/arXiv.2508.12939 , urldate =. arXiv , keywords =:2508.12939 , primaryclass =

    work page doi:10.48550/arxiv.2508.12939

  27. [27]

    arXiv: 1611.00036 , eprint =

    The. arXiv: 1611.00036 , eprint =

    Pith/arXiv arXiv

  28. [28]

    Density estimation using Real NVP

    Dinh, Laurent and. Density Estimation Using. doi:10.48550/arXiv.1605.08803 , urldate =. arXiv , keywords =:1605.08803 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1605.08803

  29. [29]

    NICE: Non-linear Independent Components Estimation

    Dinh, Laurent and Krueger, David and Bengio, Yoshua , year = 2015, month = apr, number =. doi:10.48550/arXiv.1410.8516 , urldate =. arXiv , keywords =:1410.8516 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1410.8516 2015

  30. [30]

    doi:10.48550/arXiv.2403.18514 , urldate =

    Djahnine, Aissam and Popoff, Alexandre and. doi:10.48550/arXiv.2403.18514 , urldate =. arXiv , keywords =:2403.18514 , primaryclass =

    work page doi:10.48550/arxiv.2403.18514

  31. [31]

    MNRAS , volume =

    Bayesian Inference of Initial Conditions from Non-Linear Cosmic Structures Using Field-Level Emulators , author =. MNRAS , volume =. doi:10.1093/mnras/stae2429 , abstract =. https://academic.oup.com/mnras/article-pdf/535/2/1258/60481350/stae2429.pdf , pages =

    work page doi:10.1093/mnras/stae2429

  32. [32]

    Learning the

    Doeser, Ludvig and Ata, Metin and Jasche, Jens , year = 2025, month = aug, journal =. Learning the. doi:10.1093/mnras/staf1289 , urldate =. arXiv , keywords =:2502.13243 , primaryclass =

    work page doi:10.1093/mnras/staf1289 2025

  33. [33]

    Cosmology with the

    Dor. Cosmology with the. arXiv , file =:1412.4872 , doi =

    Pith/arXiv arXiv

  34. [34]

    Dor. Science. arXiv , file =:1805.05489 , urldate =

    Pith/arXiv arXiv

  35. [35]

    Duane, Simon and Kennedy, A. D. and Pendleton, Brian J. and Roweth, Duncan , year = 1987, month = sep, journal =. Hybrid. doi:10.1016/0370-2693(87)91197-X , urldate =

    work page doi:10.1016/0370-2693(87)91197-x 1987

  36. [36]

    FastPM: a new scheme for fast simulations of dark matter and halos

    Feng, Yu and Chu, Man Yat and Seljak, Uro. Monthly Notices of the Royal Astronomical Society , volume =. doi:10.1093/mnras/stw2123 , abstract =. arXiv , keywords =:1603.00476 , pages =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1093/mnras/stw2123

  37. [37]

    Exploring the posterior surface of the large scale structure reconstruction

    Exploring the Posterior Surface of the Large Scale Structure Reconstruction , author =. Journal of Cosmology and Astroparticle Physics , volume =. doi:10.1088/1475-7516/2018/07/043 , abstract =. arXiv , arxivid =:1804.09687 , issn =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/1475-7516/2018/07/043 2018

  38. [38]

    Improving Constraints on Primordial Non-

    Fl. Improving Constraints on Primordial Non-. Journal of Cosmology and Astroparticle Physics , volume =. doi:10.1088/1475-7516/2024/02/031 , urldate =

    work page doi:10.1088/1475-7516/2024/02/031 2024

  39. [39]

    doi:10.48550/arXiv.2211.12724 , urldate =

    Friedman, Roy and Hassan, Sultan , year = 2022, month = nov, number =. doi:10.48550/arXiv.2211.12724 , urldate =. arXiv , keywords =:2211.12724 , primaryclass =

    work page doi:10.48550/arxiv.2211.12724 2022

  40. [40]

    Neural Networks , volume =

    On the Approximate Realization of Continuous Mappings by Neural Networks , author =. Neural Networks , volume =. doi:10.1016/0893-6080(89)90003-8 , urldate =

    work page doi:10.1016/0893-6080(89)90003-8

  41. [41]

    eLife , volume =

    Training Deep Neural Density Estimators to Identify Mechanistic Models of Neural Dynamics , author =. eLife , volume =. doi:10.7554/eLife.56261 , urldate =

    work page doi:10.7554/elife.56261

  42. [42]

    Goodfellow, Ian and Bengio, Yoshua and Courville, Aaron , year = 2016, publisher =. Deep

    2016

  43. [43]

    Automatic Posterior Transformation for Likelihood-Free Inference

    Greenberg, David S. and Nonnenmacher, Marcel and Macke, Jakob H. , year = 2019, month = may, number =. Automatic. doi:10.48550/arXiv.1905.07488 , urldate =. arXiv , keywords =:1905.07488 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1905.07488 2019

  44. [44]

    Proceedings of the National Academy of Sciences of the United States of America , volume =

    Learning to Predict the Cosmological Structure Formation , author =. Proceedings of the National Academy of Sciences of the United States of America , volume =. doi:10.1073/pnas.1821458116 , abstract =. arXiv , arxivid =:1811.06533 , pages =

    work page doi:10.1073/pnas.1821458116

  45. [45]

    Likelihood-Free

    Hermans, Joeri and Begy, Volodimir and Louppe, Gilles , year = 2020, month = jun, number =. Likelihood-Free. doi:10.48550/arXiv.1903.04057 , urldate =. arXiv , keywords =:1903.04057 , primaryclass =

    work page doi:10.48550/arxiv.1903.04057 2020

  46. [46]

    Denoising Diffusion Probabilistic Models

    Ho, Jonathan and Jain, Ajay and Abbeel, Pieter , year = 2020, month = dec, number =. Denoising. doi:10.48550/arXiv.2006.11239 , urldate =. arXiv , keywords =:2006.11239 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2006.11239 2020

  47. [47]

    J., Chartier , N., et al

    Ho, Matthew and Bartlett, Deaglan J. and Chartier, Nicolas and. The Open Journal of Astrophysics , volume =. doi:10.33232/001c.120559 , urldate =. arXiv , keywords =:2402.05137 , primaryclass =

    work page doi:10.33232/001c.120559

  48. [48]

    Neural Networks , volume =

    Multilayer Feedforward Networks Are Universal Approximators , author =. Neural Networks , volume =. doi:10.1016/0893-6080(89)90020-8 , urldate =

    work page doi:10.1016/0893-6080(89)90020-8

  49. [49]

    doi:10.3847/1538-4357/ab4d4c , abstract =

    Horowitz, Benjamin and Lee, Khee-Gan and White, Martin and Krolewski, Alex and Ata, Metin , year = 2019, journal =. doi:10.3847/1538-4357/ab4d4c , abstract =. arXiv , arxivid =:1903.09049v3 , pages =

    work page doi:10.3847/1538-4357/ab4d4c 2019

  50. [50]

    Monthly Notices of the Royal Astronomical Society , volume =

    Joint Cosmic Density Reconstruction from Photometric and Spectroscopic Samples , author =. Monthly Notices of the Royal Astronomical Society , volume =. arXiv , arxivid =:2311.18738 , pages =

    arXiv

  51. [51]

    doi:10.48550/arXiv.2510.19224 , urldate =

    Horowitz, Benjamin and. doi:10.48550/arXiv.2510.19224 , urldate =. arXiv , keywords =:2510.19224 , primaryclass =

    work page doi:10.48550/arxiv.2510.19224

  52. [52]

    , year = 2025, month = dec, number =

    Horowitz, Benjamin and Bayer, Adrian E. , year = 2025, month = dec, number =. doi:10.48550/arXiv.2510.26851 , urldate =. arXiv , keywords =:2510.26851 , primaryclass =

    work page doi:10.48550/arxiv.2510.26851 2025

  53. [53]

    LSST: from Science Drivers to Reference Design and Anticipated Data Products

    Ivezi. The Astrophysical Journal , volume =. doi:10.3847/1538-4357/ab042c , abstract =. arXiv , arxivid =:0805.2366 , pages =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.3847/1538-4357/ab042c

  54. [54]

    arXiv: 2206.04594 , eprint =

    Field Level Neural Network Emulator for Cosmological N-Body Simulations , author =. arXiv: 2206.04594 , eprint =

    arXiv

  55. [55]

    Field-Level

    Jamieson, Drew and Li, Yin and. Field-Level. JCAP , volume =. arXiv , file =:2408.07699 , doi =

    arXiv

  56. [56]

    Bayesian physical reconstruction of initial conditions from large scale structure surveys

    Bayesian Physical Reconstruction of Initial Conditions from Large-Scale Structure Surveys , author =. MNRAS , volume =. doi:10.1093/mnras/stt449 , urldate =. arXiv , keywords =:1203.3639 , pages =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1093/mnras/stt449

  57. [57]

    Past and present cosmic structure in the SDSS DR7 main sample

    Jasche, J. and Leclercq, F. and Wandelt, B. D. , year = 2015, journal =. Past and Present Cosmic Structure in the. doi:10.1088/1475-7516/2015/01/036 , abstract =. arXiv , keywords =:1409.6308 , issn =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/1475-7516/2015/01/036 2015

  58. [58]

    Physical Bayesian modelling of the non-linear matter distribution: new insights into the Nearby Universe

    Jasche, J. and Lavaux, G. , year = 2019, journal =. Physical. doi:10.1051/0004-6361/201833710 , abstract =. arXiv , keywords =:1806.11117 , issn =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1051/0004-6361/201833710 2019

  59. [59]

    Jiang, Heinrich , year = 2017, month = jul, pages =. Uniform. Proceedings of the 34th

    2017

  60. [60]

    arXiv: 2303.13056 , eprint =

    Predicting the Initial Conditions of the Universe Using Deep Learning , author =. arXiv: 2303.13056 , eprint =

    arXiv

  61. [61]

    doi:10.48550/arXiv.2507.11842 , urldate =

    Kannan, Sidharth and Qiu, Tian and. doi:10.48550/arXiv.2507.11842 , urldate =. arXiv , keywords =:2507.11842 , primaryclass =

    work page doi:10.48550/arxiv.2507.11842

  62. [62]

    Analyzing and

    Karras, Tero and Laine, Samuli and Aittala, Miika and Hellsten, Janne and Lehtinen, Jaakko and Aila, Timo , year = 2020, month = mar, number =. Analyzing and. doi:10.48550/arXiv.1912.04958 , urldate =. arXiv , keywords =:1912.04958 , primaryclass =

    work page doi:10.48550/arxiv.1912.04958 2020

  63. [64]

    The Astrophysical Journal , volume =

    Kaushal, Neerav and. The Astrophysical Journal , volume =. doi:10.3847/1538-4357/ac5c4a , abstract =

    work page doi:10.3847/1538-4357/ac5c4a

  64. [65]

    Glow: Generative Flow with Invertible 1x1 Convolutions

    Kingma, Diederik P. and Dhariwal, Prafulla , year = 2018, month = jul, number =. Glow:. doi:10.48550/arXiv.1807.03039 , urldate =. arXiv , keywords =:1807.03039 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1807.03039 2018

  65. [66]

    and Salimans, Tim and Poole, Ben and Ho, Jonathan , year = 2023, month = apr, number =

    Kingma, Diederik P. and Salimans, Tim and Poole, Ben and Ho, Jonathan , year = 2023, month = apr, number =. Variational. doi:10.48550/arXiv.2107.00630 , urldate =. arXiv , keywords =:2107.00630 , primaryclass =

    work page doi:10.48550/arxiv.2107.00630 2023

  66. [67]

    Monthly Notices of the Royal Astronomical Society: Letters , volume =

    The Initial Conditions of the Universe from Constrained Simulations , author =. Monthly Notices of the Royal Astronomical Society: Letters , volume =. doi:10.1093/mnrasl/sls029 , urldate =

    work page doi:10.1093/mnrasl/sls029

  67. [68]

    and Johnson, Hayden R

    Krouglova, Anastasia N. and Johnson, Hayden R. and Confavreux, Basile and Deistler, Michael and Gon. Multifidelity. doi:10.48550/arXiv.2502.08416 , urldate =. arXiv , keywords =:2502.08416 , primaryclass =

    work page doi:10.48550/arxiv.2502.08416

  68. [69]

    arXiv: 2207.05509 , eprint =

    Hybrid Physical-Neural Odes for Fast n-Body Simulations , author =. arXiv: 2207.05509 , eprint =

    arXiv

  69. [70]

    arXiv: 1110.3193 , eprint =

    Euclid Definition Study Report , author =. arXiv: 1110.3193 , eprint =

    Pith/arXiv arXiv

  70. [71]

    Unmasking the Masked Universe: the 2M++ catalogue through Bayesian eyes

    Lavaux, Guilhem and Jasche, Jens , year = 2016, journal =. Unmasking the Masked Universe:. doi:10.1093/mnras/stv2499 , abstract =. arXiv , arxivid =:1509.05040 , pages =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1093/mnras/stv2499 2016

  71. [72]

    Bayesian analysis of the dynamic cosmic web in the SDSS galaxy survey

    Leclercq, Florent and Jasche, Jens and Wandelt, Benjamin , year = 2015, journal =. Bayesian Analysis of the Dynamic Cosmic Web in the. doi:10.1088/1475-7516/2015/06/015 , abstract =. arXiv , arxivid =:1502.02690 , pages =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.1088/1475-7516/2015/06/015 2015

  72. [73]

    Posterior

    Legin, Ronan and Ho, Matthew and Lemos, Pablo and. Posterior. MNRAS , volume =. doi:10.1093/mnrasl/slad152 , urldate =. arXiv , keywords =:2304.03788 , pages =

    work page doi:10.1093/mnrasl/slad152

  73. [74]

    Sampling-

    Lemos, Pablo and Coogan, Adam and Hezaveh, Yashar and. Sampling-. doi:10.48550/arXiv.2302.03026 , urldate =. arXiv , keywords =:2302.03026 , primaryclass =

    work page doi:10.48550/arxiv.2302.03026

  74. [75]

    Lipman, Yaron and Chen, Ricky T. Q. and. Flow. doi:10.48550/arXiv.2210.02747 , urldate =. arXiv , keywords =:2210.02747 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2210.02747

  75. [76]

    arXiv , arxivid =:2310.19910 , abstract =

    Bayesian Simulation-Based Inference for Cosmological Initial Conditions , author =. arXiv , arxivid =:2310.19910 , abstract =

    arXiv

  76. [77]

    DISCO-DJ II: a differentiable particle-mesh code for cosmology

    List, Florian and Hahn, Oliver and Fl. doi:10.48550/arXiv.2510.05206 , urldate =. arXiv , keywords =:2510.05206 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2510.05206

  77. [78]

    arXiv: 1211.0310 , eprint =

    Large Synoptic Survey Telescope:. arXiv: 1211.0310 , eprint =

    Pith/arXiv arXiv

  78. [79]

    arXiv , arxivid =:0912.0201 , abstract =

    arXiv: 0912.0201 , number =. arXiv , arxivid =:0912.0201 , abstract =

    Pith/arXiv arXiv

  79. [80]

    Benchmarking

    Lueckmann, Jan-Matthis and Boelts, Jan and Greenberg, David S and Gon. Benchmarking

  80. [81]

    Lu, Zhou and Pu, Hongming and Wang, Feicheng and Hu, Zhiqiang and Wang, Liwei , year = 2017, month = nov, number =. The. doi:10.48550/arXiv.1709.02540 , urldate =. arXiv , keywords =:1709.02540 , primaryclass =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1709.02540 2017

Showing first 80 references.